CN114817976B - Sensor data protection method, system, computer equipment and intelligent terminal - Google Patents

Abstract

The invention belongs to the technical field of information data security, and discloses a sensor data protection method, a system, computer equipment and an intelligent terminal, wherein a random walk algorithm and a training method for generating an countermeasure network are adopted, a user does not need to define a specific action sequence or consume local computing resources to perform data synthesis, the user only needs to define the proportion of each action before use, then predefined data is transmitted to a cloud server, the cloud server completes action sequence construction and multi-sensor simulation data generation, a simulation data set formed by combining the simulation data with the action sequence is transmitted to a request initiator, the request initiator decomposes the simulation data set, and a Hook method is utilized to replace local sensor interface data, so that the effect of complete anonymization on a multi-sensor of mobile equipment is finally achieved.

Description

Sensor data protection method, system, computer equipment and intelligent terminal

Technical Field

The present invention belongs to the technical field of information data security, and in particular relates to a sensor data protection method, system, computer equipment and intelligent terminal.

Background technique

With the development of mobile Internet, the integration of new technologies such as smart terminals and location services has led to unprecedented development of mobile applications and services. Sensors embedded in personal smart devices have brought users a convenient user experience in personalized mobile applications. Data generated by sensors such as accelerometers, gyroscopes, and magnetometers can be used to monitor users' physical activities, interactions, and emotions. Applications installed on wearable devices can obtain raw sensor data and make inferences for tasks such as gesture recognition or activity recognition. Existing research shows that motion sensors can be used as a medium for side channel attacks to steal users' sensitive inputs, obtain users' motion status, and identify and track characteristic devices. More importantly, obtaining sensor data does not require users to grant permissions, which makes it easy to implement privacy data inference based on motion sensors and is extremely concealed.

At present, the privacy protection strategy of sensor data uses false random data or distorted data such as resampling to provide applications and similar methods, which will inevitably reduce the precision and accuracy of sensor data in usability identification, such as action recognition and step counting, so that the provided data is significantly different from the real data, and false random data is easily identified by service providers, which may cause the application to crash and thus fail to provide services to users. In addition, the current defense strategies are all carried out under the premise of ensuring usability, without considering the privacy protection of the entire life cycle. If users also need to protect user background knowledge such as motion information, the current defense strategies cannot effectively protect it. The current simulation data generation method is limited to using generative adversarial networks to solve the generation problem, and cannot effectively solve the problem of small generated data space; when the number of iterations reaches a certain number, simulation data with high similarity will appear. In addition, the current simulation data generation method only targets a single sensor when generating data. When the application requires joint judgment of multiple sensors, it cannot complete the coordination of multiple sensors, resulting in distortion of the simulation data of the target behavior.

There are some drawbacks in the current defense measures. Providing random false data or distorted data after resampling to the application will reduce precision and accuracy and bring large errors. Providing fuzzy data processed by the model to the application will improve availability, but the model processing time is long, which cannot meet the timeliness of sensor data and cannot be truly available. For the generation of simulated data, the existing data generation scheme based on generative adversarial networks has the problem of mode collapse and is only feasible in small batches and small ranges. The existing technology does not take into account complete anonymity. When the attacker obtains certain background knowledge, it will affect the protection of user privacy to a certain extent. Specifically, for example, the patent "Android terminal sensor information protection method based on differential privacy", patent number CN201810257632.4, this method chooses to add specific Laplace noise to the real data. Due to the use of real data, compared with the full obfuscation method, this scheme will still leak some background information. The patent "A sensor data generation model and method based on conditional generative adversarial network", patent number is CN202110312274.4. This solution model can only solve the generation problem. It can generate specific sensor data for specific actions to achieve the effect of realistic data generation. However, it cannot solve the mode collapse problem caused by the repeated use of the generative adversarial network model. There is a problem of large-scale data duplication, and the model only generates data for a single sensor, and cannot complete the coordination of multiple sensors.

Through the above analysis, the problems and defects of the prior art are as follows:

(1) The current defense strategies all provide false random data or distorted data such as resampling to the application, which will inevitably reduce the precision and accuracy of its availability identification, making the provided data significantly different from the real data. Moreover, false random data can be easily identified by the service provider, which may cause the application to crash. Moreover, the current defense strategies are all carried out under the premise of ensuring availability, without considering the privacy protection of the entire life cycle. If the user also needs to protect the user background knowledge such as movement information, the current defense strategies cannot provide effective protection.

(2) The current simulation data generation method is limited to using generative adversarial networks to solve the generation problem, and cannot effectively solve the problem of small generated data space; when the number of iterations reaches a certain number, simulation data with high similarity will appear. In addition, the current simulation data generation method only targets a single sensor when generating data. When the application requires multi-sensor joint judgment, it cannot complete the coordination of multiple sensors, resulting in distortion of the simulation data of the target behavior.

(3) Existing technologies do not take complete anonymity into account. When the attacker obtains certain background knowledge, it will affect the protection of user privacy to a certain extent.

The difficulty of solving the above problems and defects is as follows: sensor data replacement based on Android terminals requires replacing the code of the Android system framework layer during the operation of the mobile device, which is quite difficult; for the privacy protection of sensor data throughout the entire life cycle with complete anonymity, for the action proportion and transition probability specified by the user, an action sequence that conforms to the predefined distribution and transition probability is generated, and multiple sensors are combined to generate simulated data that conforms to the specified action classification.

The significance of solving the above problems and defects is as follows: By introducing the action sequence generation method based on the Monte Carlo method, it is suitable for constructing false action behavior sequences. By introducing the data generation method based on the time-series generative adversarial network and the filter combination, it is suitable for constructing simulated data that conforms to the specified action classification. This method achieves full anonymous privacy protection by replacing sensor data at all times and in all directions. It greatly improves the security of Android terminal sensor information and has important theoretical value and practical significance for the privacy protection of future mobile terminals.

Summary of the invention

In view of the problems existing in the prior art, the present invention provides a sensor data protection method, system, computer equipment and intelligent terminal.

The present invention is implemented as follows: a sensor data protection method, which uses a random walk algorithm and a training method of a generative adversarial network to perform data synthesis, and defines the proportion of each action before use; the predefined data is handed over to a cloud server, and the cloud server completes the action sequence construction and multi-sensor simulated data generation, and a simulated data set formed by combining the simulated data with the action sequence is handed over to a request initiator; the request initiator decomposes the simulated data set, and uses the Hook method to replace the local sensor interface data, so as to achieve complete anonymization on multiple sensors of a mobile device.

Furthermore, the sensor data protection method replaces the sensor data by using the sensor simulated data sequence at all times and in all directions.

Furthermore, the multi-sensor simulation data generation adopts a random walk algorithm based on the Markov chain Monte Carlo method, and by introducing the transition probability between actions as a suggested distribution for constructing the Markov matrix, the receiving distribution in the process of constructing the Markov chain is improved, so that after the random walk algorithm is completed, a behavior action sequence that conforms to the predefined distribution can be generated;

The multi-sensor simulated data generation adopts a time series-based generative adversarial network model combined with a filter combination method based on Bayesian optimization. The time series generative adversarial network generates data sensor data that conforms to a predefined classification, introduces a filter combination method and uses Bayesian optimization to search for filter combination parameters that meet the requirements.

Further, the sensor data protection method comprises the following steps:

The first step is to initialize the system. The user inputs the action distribution, including the proportion of standing, walking, running, sitting, lying, and going up and down stairs. The transfer matrix is constructed by the predefined action transition probability and the action distribution input by the user, and multiple rounds of iterations are performed to verify whether a stable distribution is achieved, providing feasibility support for the subsequent generation of action sequences; the state transfer matrix P _ij = p(i, j)i, j∈S is calculated by the formula p(x, x′) = q(x, x′) α(x, x′), where S represents all behavior action states. By initializing the vector λ ₀ = {1, 0, 0, 0, 0}, substituting it into the formula λ _t = λ _t-1 P, where P represents the state transfer matrix, the distribution after t rounds of iteration is obtained;

The second step is to construct a simulated action sequence. The proposed distribution and acceptance distribution of the constructed transfer matrix are used in combination with the random walk algorithm to generate a behavior action sequence that conforms to the predefined action distribution, providing data support for the subsequent sensor simulation data arrangement rules; the random walk algorithm based on the Markov chain Monte Carlo method is used to generate the action sequence, and the acceptance distribution is directly used in the random walk algorithm:

Where p(x′) represents the distribution of state x′, and p(x) represents the distribution of state x;

The third step is to generate sensor simulated data. The adversarial network model is trained with real data in advance, so that the accuracy of the simulated data generated by the model in the action recognition task reaches more than 90%. For each action, multiple sets of data are generated as a buffer to provide the original data template for the subsequent simulated data space expansion task.

The fourth step is to expand the simulated data space, take out the data of each action in the buffer, combine them according to the filter combination rules, and use the Bayesian optimization algorithm to select multiple parameters that can achieve local optimality;

The fifth step is data combination and replacement. The simulated data is filled in according to the behavioral action sequence generated by the sensor simulated data. The sensor data distribution interface is hooked at the bottom layer of the mobile device to replace the batch of sensor data. The privacy and security of sensor data are protected from the mobile terminal data release link.

Furthermore, the second step of generating a simulated action sequence is as follows: a random walk algorithm based on the Markov chain Monte Carlo method is used to construct a state transfer matrix according to the action ratio preset by the user to generate a false action sequence; the Monte Carlo method is used to construct a Markov transfer matrix using the transfer kernel formula:

p(x, x′)=q(x, x′)α(x, x′);

In the formula, q(x, x′) is called the proposed distribution, and α(x, x′) is called the received distribution. The proposed distribution is symmetric, and the received distribution is:

In the formula, p(x′) represents the proportion of state x′, and p(x) represents the proportion of state x; the recommended distribution is the transition probability from state x to state x′, satisfying In the formula, X represents the set of states adjacent to state x and includes state x;

The value function of the adversarial network model in the third step is:

Wherein, the first part on the right side of the formula equal sign represents the expectation of the discriminator trained on the real data represented by the high-dimensional latent space, and the second part represents the expectation of the discriminator trained on the synthetic data of the high-order latent space synthesized by the generator; where G represents the generator network, D represents the discriminator network, E represents the expectation, x~p _data (x) represents the real data sampled from the real data set, log represents the logarithmic function, x represents the real data, X represents the real data represented by the high-dimensional latent space, z~p _z (z) represents the random noise vector sampled from the normal distribution, and z represents the random noise vector;

The embedding recovery loss calculation uses the following formula to calculate the difference between the original data and the data processed by the embedding function module and the recovery function module:

In the formula, l _R represents the difference between the original data and the restored data, E represents the mathematical expectation, x _t represents the original data, Represents the data mapped from the original space to the latent space, and from the latent space to the original space, ||...|| ₂ represents the L2 norm;

The binary judgment module uses the following loss function to calculate the difference between real data and synthetic data during training:

In the formula, l _U represents the cross entropy function between real data and synthetic data, y _t represents real data, Represents synthetic data.

Furthermore, the fourth step of expanding the simulated data space adopts a filter combination method to achieve the expansion of the data space, and at the same time adopts a Bayesian optimization method to find various parameters of the filter combination;

Among them, the filter combination method is to combine the original data and the filtered data according to the formula after the frequency and combination ratio are cut off according to the simulated data generated by the generative adversarial network:

f ₁ (x ₁ , x ₂ , x ₃ )=x ₁ *filter (x ₂ , data)+x ₃ *data;

In the formula, the first part on the right side of the equal sign represents a certain proportion of the original data, and the second part represents a certain proportion of the filtered data; _x1 represents the proportion of the filtered data in the combined data, _x2 represents the cutoff frequency of the filter, data represents the original data, filter( _x2 , data) represents the filtered data, and _x3 represents the proportion of the original data in the combined data; the left side of the equal sign represents the result after the filtered combination;

Use Bayesian optimization algorithm to find parameters (x ₁ , x ₂ , x ₃ ), including optimizing expressions, fitting models, and collecting functions;

Determine the optimization expression, use the Gaussian process as the fitting model, and the probability boost function as the acquisition function:

f ₂ (x ₁ , x ₂ , x ₃ )=dtw (f ₁ (x ₁ , x ₂ , x ₃ ), data);

In the formula, the right side of the formula equal sign represents the distance between the filtered combined data and the original data, and the left side of the formula equal sign represents the specific value of the distance; where dtw represents the dynamic time adjustment distance calculation function, f ₁ (x ₁ , x ₂ , x ₃ ) represents the filtered combined data, and data represents the original data;

The interception and replacement of sensor data in the fifth step: implement the sensor monitoring module through Hook, and perform bottom-level interception and replacement on the sensor data transmission interface; find the module class android.hardware.SystemSensorManager that controls the distribution of sensor data in the Android8.0 system source code, find the specific sensor processing subclass SensorEventQueue in the module class, and the distribution function dispatchSensorEvent therein; Hook the dispatchSensorEvent method under SystemSensorManager in the system service process, load the pre-compiled replacement function module, and replace the sensor interface with synthetic data.

Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the sensor data protection method.

Another object of the present invention is to provide an information data processing terminal, wherein the information data processing terminal is used to implement the sensor data protection method.

Another object of the present invention is to provide a sensor data protection system for implementing the sensor data protection method, the sensor data protection system comprising:

The system initialization module is used to realize the user input action distribution, build a transfer matrix through the predefined action transition probability and the user input action distribution, and perform multiple rounds of iterations to verify whether a stable distribution can be achieved, providing feasibility support for the subsequent generation of action sequences;

The simulated action sequence construction module is used to use the proposed distribution and the accepted distribution of the constructed transfer matrix, combined with the random walk algorithm to generate a behavior action sequence that conforms to the predefined action distribution, providing data support for the subsequent sensor simulated data arrangement rules;

The sensor simulation data generation module is used to pre-train the generative adversarial network model with real data, so that the accuracy of the simulation data generated by the model in the action recognition task reaches more than 90%. For each action, multiple sets of data are generated as a buffer to provide the original data template for the subsequent simulation data space expansion task;

Expand the simulated data space module to extract the data of each action in the buffer, combine them according to the filter combination rules, and use the Bayesian optimization algorithm to select multiple parameters that can achieve local optimality. Since data with large differences from the original data but the same classification can be generated, this appropriately solves the problem of mode collapse that may exist in the generative adversarial network;

The data combination and replacement module is used to generate simulated data according to the behavioral action sequence generated by the sensor simulated data, hook the sensor data distribution interface at the bottom layer of the mobile device, replace the batch of sensor data, and protect the privacy and security of sensor data from the mobile terminal data release link.

Furthermore, the sensor data protection system further includes: a generator, a discriminator;

The generator includes an embedding function module, a recovery function module, an embedding recovery loss calculation module, a multi-scale circulation module, and a timing function module;

The embedding function module is used to map data from a low dimension in the original space to a high dimension in the latent space; the recovery function module is connected to the embedding function module and is used to accurately restore data from a high-dimensional latent space to a low-dimensional real space; the embedding recovery loss calculation module is used to calculate the difference between the real data and the original data after being processed by the embedding function module and the recovery function module, and is used to repeatedly train the embedding function module and the recovery function module so that the original data can be accurately expressed in the high-dimensional space; the multi-scale cycle module is used to learn the time domain characteristics of each dimension of the multi-sensor and the correlation between the time domain characteristics of each dimension; the timing function module is used to better represent the synthetic data output by the generator in the high-dimensional latent space during the adversarial training process;

The discriminator includes a binary judgment function module and a similarity calculation module; the binary judgment function module is used to distinguish real data from synthetic data during adversarial training; the similarity calculation module is connected to the binary judgment function module and is used to calculate the cosine similarity between the synthetic data and the real data in the low-dimensional original space;

The embedding function module and the restoration function module are both composed of a multi-scale recurrent neural network and a fully connected network layer. The multi-scale recurrent neural network is composed of one-dimensional recurrent neural network layers of different sizes. The output of each node in the last layer of the multi-scale recurrent neural network is used as the input of the fully connected layer.

The timing function module includes a fully connected network and a GRU network;

The embedding recovery loss calculation module uses the following formula to calculate the difference between the original data and the data processed by the embedding function module and the recovery function module;

The binary judgment module uses the following loss function to calculate the difference between real data and synthetic data during training.

Combining all the above-mentioned technical solutions, the advantages and positive effects of the present invention are as follows: the sensor data replacement of the Android platform of the present invention is combined with the simulation data generation method, which improves the defect of poor real-time performance of the existing solution, and protects the privacy security of the sensor data throughout the life cycle from the data generation link of the mobile terminal, while preventing malicious theft and analysis of user privacy by third parties on the application server side.

The present invention generates a behavioral action sequence that conforms to a predefined distribution and a predefined transition probability through a random walk algorithm based on the Markov chain Monte Carlo method; generates sensor simulated data that conforms to predefined classifications through a sensor simulated data generation method based on a time series generative adversarial network and a simulated data space expansion method based on filter combination and Bayesian optimization, and there are obvious differences between the simulated data; and by replacing mobile terminal sensor data at all times and in all directions, a user privacy protection effect that is completely anonymized throughout the entire life cycle can be achieved for mobile device sensors.

The present invention combines the interception and replacement of sensor data on the Android platform with an Android terminal sensor data protection strategy based on multi-sensor simulated data replacement, which not only protects the privacy of sensor data from the data release link of the mobile terminal, but also effectively prevents the malicious inference of user privacy by the attacker on the server side, thereby preventing the theft of user privacy. The present invention applies statistical learning methods and deep learning methods to the privacy protection of sensor data on Android mobile terminals, which can eliminate the attacker's ability to infer user privacy. Even if the attacker collects the user's sensor data for a long time, it will not affect the security of privacy protection. The present invention proposes to generate adversarial networks to generate simulated data, so that the simulated data can maintain accuracy similar to that of real data when facing inferences such as action classification. The present invention adopts a filter combination and a Bayesian optimization algorithm to realize the expansion of the simulated data space. The filter combination increases the gap with the original data in the time domain while ensuring the frequency domain characteristics, and more conveniently increases the space of simulated data.

The present invention can ensure user privacy security on the mobile terminal and the server terminal, while repeatedly using the generative adversarial network model and dynamically adjusting the filtering parameters, so as to ensure the availability recognition accuracy and reduce the repetition rate as much as possible, achieve a better obfuscation effect on the server side, and be insensitive to the background information possessed by the attacker; the low-frequency resampling technology cannot ensure the security of the mobile terminal data because it uses real data, and can ensure that the accuracy of the server's inference of user privacy is reduced, but it cannot achieve complete obfuscation; the sensor data is intercepted and replaced completely with random data, which can ensure the security of the mobile terminal, but it is easy to be identified as an abnormal user by the server and terminate the normal function service.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a sensor data protection method provided by an embodiment of the present invention.

FIG. 2 is a schematic diagram of the structure of a sensor data protection system provided in an embodiment of the present invention.

FIG3 is a flow chart of an implementation of a sensor data protection method provided in an embodiment of the present invention.

FIG. 4 is a schematic diagram of a sensor data protection system provided in an embodiment of the present invention.

FIG5 is a flow chart of an Android terminal sensor data protection system based on multi-sensor simulated data replacement provided in an embodiment of the present invention.

FIG. 6 is a schematic diagram of common action ratios and transition probabilities provided by an embodiment of the present invention.

FIG. 7 is a schematic diagram of the steady distribution convergence of the random walk algorithm provided by an embodiment of the present invention.

FIG8 is a schematic diagram of a sensor simulation data generation model architecture of a generative adversarial network provided in an embodiment of the present invention.

FIG. 9 is a curve of real data and simulated data provided by an embodiment of the present invention, and the behavior category is a running diagram.

FIG. 10 is a curve of real data and simulated data provided by an embodiment of the present invention, and the behavior category is a schematic diagram of going downstairs.

FIG. 11 is a curve of real data and simulated data provided by an embodiment of the present invention, and the behavior category is a walking diagram.

FIG12 is a comparison between the low-frequency filter combination provided by an embodiment of the present invention and the original data, the behavior is climbing stairs, and the compared data is a schematic diagram of the accelerometer X-axis data.

FIG13 is a comparison between the high-frequency filter combination provided by an embodiment of the present invention and the original data. The behavior is climbing stairs, and the compared data is a schematic diagram of the accelerometer X-axis data.

In the figure: 1. System initialization module; 2. Simulated action sequence construction module; 3. Sensor simulated data generation module; 4. Expanded simulated data space module; 5. Data combination and replacement module; 100. Subsystem; 101. Generator; 102. Discriminator; 1011. Automatic encoder and decoder; 1012. Embedding recovery loss calculation module; 1013. Multi-scale cycle module; 1014. Timing function module; 1021. Binary function discriminator; 1022. Similarity calculation module.

Detailed ways

In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

In view of the problems existing in the prior art, the present invention provides a sensor data protection method, system, computer device and intelligent terminal. The present invention is described in detail below with reference to the accompanying drawings.

As shown in FIG1 , the sensor data protection method provided by the present invention comprises the following steps:

S101, system initialization, user input action distribution, construct a transfer matrix through predefined action transition probability and user input action distribution, and perform multiple rounds of iterations to verify whether a stable distribution can be achieved, so as to provide feasibility support for subsequent generation of action sequences;

S102, simulated action sequence construction, using the proposed distribution and accepted distribution of the constructed transfer matrix, combined with the random walk algorithm to generate a behavior action sequence that conforms to the predefined action distribution, providing data support for the subsequent sensor simulated data arrangement rules;

S103, sensor simulation data generation, using real data to train the generative adversarial network model in advance, so that the accuracy of the simulation data generated by the model in the action recognition task reaches more than 90%, and multiple sets of data are generated for each action as a buffer to provide the original data template for the subsequent simulation data space expansion task;

S104, expanding the simulated data space, taking out the data of each action in the buffer, combining them according to the filter combination rule, and using the Bayesian optimization algorithm to select multiple parameters that can achieve local optimality. Since data with large differences from the original data and the same classification can be generated, this appropriately solves the problem of mode collapse that may exist in the generative adversarial network;

S105, data combination and replacement, fill in the simulated data according to the behavioral action sequence generated by the sensor simulated data, hook the sensor data distribution interface at the bottom layer of the mobile device, replace the batch of sensor data, and protect the privacy and security of sensor data from the mobile terminal data release link.

The sensor data protection method provided by the present invention may be implemented by ordinary technicians in the industry using other steps. The sensor data protection method provided by the present invention in FIG. 1 is only a specific embodiment.

As shown in FIG. 2 and FIG. 4 , the sensor data protection system provided by the present invention includes:

System initialization module 1 is used to realize the user input action distribution, construct a transfer matrix through the predefined action transition probability and the user input action distribution, and perform multiple rounds of iterations to verify whether a stable distribution can be achieved, so as to provide feasibility support for the subsequent generation of action sequences;

The simulated action sequence construction module 2 is used to use the proposed distribution and the accepted distribution of the constructed transfer matrix, combined with the random walk algorithm to generate a behavior action sequence that conforms to the predefined action distribution, and provide data support for the subsequent sensor simulated data arrangement rules;

The sensor simulation data generation module 3 is used to train the generative adversarial network model with real data in advance, so that the accuracy of the simulation data generated by the model in the action recognition task reaches more than 90%. For each action, multiple sets of data are generated as a buffer to provide the original data template for the subsequent simulation data space expansion task;

Expanding the simulated data space module 4 is used to take out the data of each action in the buffer, combine them according to the filter combination rule, and use the Bayesian optimization algorithm to select multiple parameters that can achieve local optimality. Since data with large differences from the original data and the same classification can be generated, this appropriately solves the problem of mode collapse that may exist in the generative adversarial network;

The data combination and replacement module 5 is used to fill in the simulated data according to the behavioral action sequence generated by the sensor simulated data, hook the sensor data distribution interface at the bottom layer of the mobile device, replace the batch of sensor data, and protect the privacy and security of the sensor data from the mobile terminal data release link.

The technical solution of the present invention is further described below in conjunction with specific embodiments.

Embodiment 1:

The sensor data protection method provided by the present invention replaces sensor data with a sensor simulated data sequence at all times and in all directions, thereby achieving a privacy protection effect in the entire life cycle and with full anonymity. Improvements are made to the sensor data privacy protection method based on multi-sensor simulated data replacement and the multi-sensor simulated data generation method.

The sensor data privacy protection method of the present invention analyzes the differential privacy protection method of existing privacy protection, and uses highly simulated multi-sensor data to replace the mobile device sensors at all times and in all directions, targeting the limitation of the current scheme that uses real data processing to cause possible user background leakage risks. During the whole process, the server cannot obtain the user's real data. The server can perform classification tasks such as action classification on the incoming simulated data, but cannot obtain any background information and privacy data of the user.

In order to generate an action sequence that conforms to a predefined distribution, the multi-sensor simulated data generation method of the present invention designs a random walk algorithm based on the Markov chain Monte Carlo method, and introduces the transition probability between actions as a suggested distribution for constructing the Markov matrix to improve the receiving distribution in the process of constructing the Markov chain, so that after the random walk algorithm is completed, a behavioral action sequence that conforms to the predefined distribution can be generated; in order to generate multi-sensor simulated data that performs well in action classification tasks and has a low repetition rate, a time series-based generative adversarial network model combined with a filtering combination method based on Bayesian optimization is designed, and data sensor data that conforms to the predefined classification is generated through a time series generative adversarial network, a filtering combination method is introduced to increase the space of simulated data, and Bayesian optimization is used to search for filtering combination parameters that meet the requirements, while solving the mode collapse problem of the generative adversarial network.

Use the above steps to generate realistic data sequences of multiple sensors of mobile devices in real scenarios.

Embodiment 2:

The present invention combines the sensor data replacement of the Android terminal sensor data protection method based on multi-sensor simulated data replacement with the simulated data generation method of the _Android id platform, protects the privacy security of sensor data from the data generation link of the mobile terminal, and prevents the malicious theft and analysis of user privacy by a third party at the application service end; specifically, it includes the following steps:

Step 1: Action sequence generation: A random walk algorithm based on the Markov chain Monte Carlo method is used to construct a state transfer matrix according to the action ratio preset by the user to generate a false behavior action sequence; the Monte Carlo method is used to construct the Markov transfer matrix using the transfer kernel formula:

p(x, x′)=q(x, x′)α(x, x′);

In the formula, q(x, x′) is called the proposed distribution, and α(x, x′) is called the received distribution. Assuming that the proposed distribution is symmetric, the received distribution is:

In the formula, p(x′) represents the proportion of state x′, and p(x) represents the proportion of state x. The recommended distribution is the transition probability from state x to state x′, satisfying Wherein, X represents the set of states adjacent to state x and including state x.

Step 2, preliminarily generate simulated data: use a time series generative adversarial network to generate multi-sensor simulated data, including: a generator, a discriminator, wherein the generator includes an embedding function module, a recovery function module, an embedding recovery loss calculation module, a multi-scale cycle module, and a time series function module; the embedding function module is used to map data from a low dimension in the original space to a high dimension in the latent space; the recovery function module is connected to the embedding function module, and is used to accurately restore data from a high-dimensional latent space to a low-dimensional real space; the embedding recovery loss calculation module is used to calculate the difference between the real data after being processed by the embedding function module and the recovery function module and the original data, and is used to repeatedly train the embedding function module and the recovery function module so that the original data can be accurately expressed in a high-dimensional space; the multi-scale cycle module is used to learn the time domain characteristics of each dimension of the multi-sensor and the correlation between the time domain characteristics of each dimension; the time series function module is used to better represent the synthetic data output by the generator in a high-dimensional latent space during adversarial training;

The discriminator includes a binary judgment function module and a similarity calculation module; the binary judgment function module is used to distinguish real data from synthetic data during adversarial training; the similarity calculation module is connected to the binary judgment function module and is used to calculate the cosine similarity between synthetic data and real data in a low-dimensional original space.

The value function of the model is:

Among them, the first part on the right side of the formula represents the expectation of the discriminator being trained on the real data represented by the high-dimensional latent space, and the second part represents the expectation of the discriminator being trained on the synthetic data of the high-dimensional latent space synthesized by the generator; wherein G represents the generator network, D represents the discriminator network, E represents the expectation, x~p _data (x) represents the real data sampled from the real data set, log represents the logarithmic function, x represents the real data, X represents the real data represented by the high-dimensional latent space, z~p _z (z) represents the random noise vector sampled from the normal distribution, and z represents the random noise vector.

The embedding function module and the recovery function module are both composed of a multi-scale recurrent neural network and a fully connected network layer. The multi-scale recurrent neural network is composed of one-dimensional recurrent neural network layers of different sizes. The output of each node in the last layer of the multi-scale recurrent neural network is used as the input of the fully connected layer.

The timing function module includes a fully connected network and a GRU network.

The embedding recovery loss calculation module uses the following formula to calculate the difference between the original data and the data processed by the embedding function module and the recovery function module:

In the formula, l _R represents the difference between the original data and the restored data, E represents the mathematical expectation, x _t represents the original data, represents the data mapped from the original space to the latent space and from the latent space to the original space, ||...|| ₂ represents the L2 norm.

The binary judgment module uses the following loss function to calculate the difference between real data and synthetic data during training:

In the formula, l _U represents the cross entropy function between real data and synthetic data, y _t represents real data, Represents synthetic data.

Step 3: Expand the simulated data space: Use the filter combination method to expand the data space, and use the Bayesian optimization method to find the parameters of the filter combination.

Among them, the filter combination method is to combine the original data and the filtered data according to the following formula after setting the cutoff frequency and combination ratio based on the simulated data generated by the generative adversarial network:

f1(x ₁ , x ₂ , x ₃ )=x ₁ *filter(x ₂ , data)+x ₃ *data;

In the formula, the first part on the right side of the equal sign represents a certain proportion of the original data, and the second part represents a certain proportion of the filtered data; _x1 represents the proportion of filtered data in the combined data, _x2 represents the cutoff frequency of the filter, data represents the original data, filter( _x2 , data) represents the data after filtering, and _x3 represents the proportion of the original data in the combined data; the left side of the equal sign represents the result after filtering combination.

A Bayesian optimization algorithm is used to find parameters (x ₁ , x ₂ , x ₃ ), including optimizing expressions, fitting models, and collecting functions.

Determine the following formula as the optimization expression, Gaussian process as the fitting model, and probability boost function as the acquisition function:

f ₂ (x ₁ , x ₂ , x ₃ )=dtw (f ₁ (x ₁ , x ₂ , x ₃ ), data);

In the formula, the right side of the equal sign represents the distance between the filtered combined data and the original data, and the left side of the equal sign represents the specific value of the distance; wherein dtw represents the dynamic time adjustment distance calculation function, f ₁ (x ₁ , x ₂ , x ₃ ) represents the filtered combined data, and data represents the original data.

Step 4: Sensor data interception and replacement: All applications under the Android system are hatched by the Zygote process; the executable program corresponding to the startup of the Zygote process is app_process. By replacing the system's app_process executable file and the virtual machine dynamic link library, Zygote can inject the module code when starting the application process. Implement the sensor monitoring module through Hook, and perform bottom-level interception and replacement of the sensor data transmission interface; find the module class android.hardware.SystemSensorManager that controls the distribution of sensor data in the Android8.0 system source code, find the specific sensor processing subclass SensorEventQueue in the module class, and the dispatch function dispatchSensorEvent therein; Hook the dispatchSensorEvent method under SystemSensorManager in the system service process, load the pre-compiled replacement function module, and replace the sensor interface with synthetic data.

In step 1 of the present invention, system initialization specifically includes:

(1) The mobile terminal user input includes the proportion of several actions such as standing, walking, running, sitting, lying, going up and down stairs, etc. FIG6 shows the proportion and transition probability defined in the embodiment.

(2) The state transfer matrix P _ij = p(i, j) i, j∈S is calculated by the action distribution ratio and the transition probability between actions in FIG6 through the formula p(x, x′) = q(x, x′) α(x, x′), where S represents all behavior action states. By initializing the vector λ ₀ = {1, 0, 0, 0, 0, 0}, substituting into the formula λ _t = λ _t-1 P, where P represents the state transfer matrix, the distribution at the time of t rounds of iteration is obtained. FIG7 shows the distribution convergence in the embodiment.

In step 2 of the present invention, the construction of the simulated action sequence specifically includes:

The random walk algorithm based on the Markov chain Monte Carlo method is used to generate the action sequence. The receiving distribution provided in step 1 is directly used in the random walk algorithm:

Where p(x′) represents the distribution of state x′, and p(x) represents the distribution of state x.

The following is a pseudo-code description of the method generation process.

The above introduces the process of random walk algorithm in detail.

In step three of the present invention, FIG8 shows the system structure according to an embodiment of the present invention. As can be seen from the figure, the subsystem 100 of the present invention includes a generator 101 and a discriminator 102. The goal of the generator 101 is to make full use of the potential time-domain and frequency-domain characteristics of the sensor data itself to learn the distribution characteristics of the sensor's real data, so as to generate sensor simulation data that is closer to the real distribution; the goal of the discriminator 102 is to combine the real data with the synthetic data for binary classification, strengthen the effect of the classifier during the adversarial training process, and measure the effect of the generator. Sensor simulation data generation specifically includes:

(1) Perform min-max normalization on the data in the real dataset and save the minimum and maximum values of the real dataset to prepare data for the model to generate simulated data and then restore it to the original scale;

(2) In the model training process, the automatic encoder/decoder 1011 under the generator 101 shown in FIG8 is first trained. The purpose of the automatic encoder/decoder 1011 is to accurately map data from the low-dimensional original space to the high-dimensional latent space and accurately restore the data in the high-dimensional latent space to the low-dimensional original space; the real data used for training is brought into the embedding function module, the real data in the low-dimensional original space is mapped to the high-dimensional latent space, and the high-dimensional form of the real data is brought into the recovery function module to obtain the data of the original dimension; the loss function of the embedding recovery loss calculation module 1012 under the generator 101 is:

This formula represents the loss function for training the automatic encoder-decoder, _Xt represents the original data of t batches, It represents the data after t batches of recovery, calculates the L2 norm, and ∑ represents the sum.

(3) The purpose of the time series function module under the generator 101 shown in FIG8 is to capture the characteristics of the real data in the high-dimensional latent space. The real data is processed by the embedding function module under the automatic codec 1011, and then brought into the time series function module 1014 and output. The output result is subjected to a binary cross entropy operation with the high-dimensional space result. The loss function of the time series function module 1014 is:

Wherein h _t represents the representation of real data in high-dimensional latent space at time t, g _X represents the temporal function module function, h _t-1 represents the representation of real data in high-dimensional latent space at time t-1, and z _t represents random data at time t. According to one embodiment of the present invention, the input and output dimensions of the multi-scale cycle module 1013 are as follows:

Time domain recurrent neural network input dimension (3D): [64, 128, 9];

Time domain recurrent neural network output dimension (three-dimensional): [64, 128, 64];

Time domain feature fully connected network input dimension (three-dimensional): [64, 128, 64];

Time domain feature fully connected network output dimension (three-dimensional): [64, 128, 64];

(4) The purpose of the binary function discriminator 1021 under the discriminator 102 shown in FIG8 is to distinguish between real data and synthetic data generated by the generator. The binary function discriminator 1021 needs to distinguish between real data and synthetic data generated by the generator in a high-dimensional space. The processing results of the discriminator on the real data and the processing results on the synthetic data need to satisfy the unsupervised loss function formula:

In the formula, _yt represents the processing result of the discriminator on the real data, According to one embodiment of the present invention, the input and output dimensions of the multi-scale cycle module 1013 are as follows:

Time domain recurrent neural network input dimension (3D): [64, 128, 64];

Time domain recurrent neural network output dimension (three-dimensional): [64, 128, 64];

Classification fully connected layer input dimension (3D): [64, 128, 64];

Classification fully connected layer output dimension (3D): [64, 128, 1].

(5) The purpose of the similarity calculation module 1022 under the discriminator 102 shown in Figure 8 is to verify the distribution of the synthetic data and the original data. It is necessary to calculate the similarity between the real data and the real data, and the similarity between the real data and the simulated data. If the two values are close, it means that the distribution of the simulated data is close to the distribution of the real data; and it is necessary to calculate the similarity between the simulated data and the simulated data to ensure that there is a gap between the simulated data; it is necessary to calculate the maximum similarity between the real data and the synthetic data, which allows the present invention to understand that in the most similar case, the similarity between the data generated by the present invention and the real data is very valuable for ensuring that the user's privacy is protected. If the maximum similarity between some simulated data and the original data is higher than 80%, step five needs to be used for data processing operations. Table 1 shows the cosine similarity of each action under each indicator when synthesizing 50 sets of data.

Table 1

Activity Real data to real data similarity Synthetic data to synthetic data similarity Similarity between real data and intermediate data The maximum similarity between real data and synthetic data Down the stairs 0.6790 0.2918 0.3011 0.7998 Go up the stairs 0.3711 0.1326 0.1997 0.7997 walk 0.9150 0.2230 0.1237 0.7997 running 0.2829 0.1067 0.0627 0.7801 Standing 0.4280 0.3459 0.3898 0.7991 average 0.5352 0.2200 0.2154 0.7957

The following is a pseudo-code description of the training process.

The above details the process of adversarial training.

In step 4, expanding the simulated data space specifically includes:

(1) The formula for synthesizing the simulated data is:

x _o (n) = r _a *x (n) + r _b *filter (x (n), f _t );

In the formula, the first part on the right side of the equal sign represents a certain proportion of original data in the combined data, and the second part represents a certain proportion of filtered data; _ra represents the proportion of original data, x(n) represents original data, _rb represents the proportion of filtered data, filter(n, _ft ) represents filtered data, _ft represents the cutoff frequency; the left side of the equal sign represents the result of the filter combination.

(2) The Bayesian optimization objective function is:

f ₁ (x ₁ , x ₂ , x ₃ )=x ₁ *filter (x ₂ , data)+x ₃ *data;

f ₂ (x ₁ , x ₂ , x ₃ )=dtw (f ₁ (x ₁ , x ₂ , x ₃ ), data);

In the formula, _x1 represents the ratio of filtered data, _x2 represents the cutoff frequency, _x3 represents the ratio of original data, data represents the original data, and dtw is the dynamic time warping distance calculation function to calculate the similarity of two time series, which is especially suitable for time series of different lengths and rhythms as an indicator to measure the difference between the filtered combination data and the original data. Figure 12 shows the comparison between the low-pass and high-pass filtered data and the original data of the accelerometer x-axis under the action of climbing stairs.

The following pseudocode explains the training process of step 4 in detail.

The above introduces the training process of step 4 in detail.

In step 5, data combination and replacement specifically include:

(1) After setting the start and end time and the action ratio, the mobile user will obtain a complete set of sensor data from the algorithm model and wait for real-time replacement.

(2) The present invention uses Hook to implement the interception module to replace the real-time sensor data. The Zygote process starts to monitor the class android.hardware.SystemSensorManager that distributes sensors in the system, as well as its sensor data processing subclass SensorEventQueue and the dispatch method dispatchSensorEvent, waiting for the data replacement module to operate.

(3) Combined with the acquired simulated data, the data replacement module is completed with the help of Java iterator, and the sensor data distribution interface in the Android system is encapsulated so that the system consumes a set of simulated data each time the interface is called.

It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and/or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims (9)

1. A sensor data protection method, characterized in that the sensor data protection method comprises the following steps: The first step is to initialize the system. The user inputs the action distribution, which includes the proportion of standing, walking, running, sitting, lying, and going up and down stairs. The transfer matrix is constructed by the predefined action transition probability and the action distribution input by the user, and multiple rounds of iterations are performed to verify whether a stable distribution is achieved, so as to provide feasibility support for the subsequent generation of action sequences. Calculate the state transition matrix , where Represents all behavior action states; through the initialization vector , into the formula , where Represents the state transfer matrix, and we get Distribution during round iterations; The second step is to construct a simulated action sequence. The proposed distribution and acceptance distribution of the constructed transfer matrix are used in combination with the random walk algorithm to generate a behavior action sequence that conforms to the predefined action distribution, providing data support for the subsequent sensor simulation data arrangement rules; the random walk algorithm based on the Markov chain Monte Carlo method is used to generate the action sequence, and the acceptance distribution is directly used in the random walk algorithm: ; In the formula, Indicates status Distribution, Indicates status Distribution; The third step is to generate sensor simulated data. The adversarial network model is trained with real data in advance, so that the accuracy of the simulated data generated by the model in the action recognition task reaches more than 90%. For each action, multiple sets of data are generated as a buffer to provide the original data template for the subsequent simulated data space expansion task. The fourth step is to expand the simulated data space, take out the data of each action in the buffer, combine them according to the filter combination rules, and use the Bayesian optimization algorithm to select multiple parameters that can achieve local optimality; The fifth step is data combination and replacement. The simulated data is filled in according to the behavioral action sequence generated by the sensor simulated data. The sensor data distribution interface is hooked at the bottom layer of the mobile device to replace the batch of sensor data. The privacy and security of sensor data are protected from the mobile terminal data release link. 2. The sensor data protection method according to claim 1 is characterized in that the sensor data protection method replaces the sensor data by using the sensor simulated data sequence at all times and in all directions. 3. The sensor data protection method according to claim 1 is characterized in that the multi-sensor simulation data generation adopts a random walk algorithm based on the Markov chain Monte Carlo method, and by introducing the transition probability between actions as a suggested distribution for constructing the Markov matrix, the receiving distribution in the process of constructing the Markov chain is improved, so that after the random walk algorithm is completed, a behavior action sequence that conforms to the predefined distribution can be generated; The multi-sensor simulated data generation adopts a time series-based generative adversarial network model combined with a filter combination method based on Bayesian optimization. The time series generative adversarial network generates data sensor data that conforms to a predefined classification, introduces a filter combination method and uses Bayesian optimization to search for filter combination parameters that meet the requirements. 4. The sensor data protection method according to claim 1 is characterized in that the second step of generating a simulated action sequence is: using a random walk algorithm based on the Markov chain Monte Carlo method, constructing a state transfer matrix according to the action ratio preset by the user, and realizing the generation of a false behavior action sequence; using the Monte Carlo method, constructing a Markov transfer matrix using a transfer kernel formula of: ; In the formula is called the proposal distribution, is called the receiving distribution; it is suggested that the distribution is symmetric and the receiving distribution is: ; In the formula, Indicates status The proportion of Indicates status The recommended distribution is from state To status The transition probability satisfies , where Representation and Status A set of adjacent states, including the state ; The value function of the adversarial network model in the third step is: ; The first part on the right side of the formula represents the expectation of the discriminator trained on real data represented by the high-dimensional latent space, and the second part represents the expectation of the discriminator trained on synthetic data of the high-dimensional latent space synthesized by the generator; represents the generator network, represents the discriminator network, Express expectations, represents the real data sampled from the real dataset, represents the logarithmic function, Represents real data, represents the real data represented in a high-dimensional latent space, represents a random noise vector sampled from a normal distribution, represents a random noise vector; The embedding recovery loss calculation uses the following formula to calculate the difference between the original data and the data processed by the embedding function module and the recovery function module: ; In the formula, Indicates the difference between the original data and the restored data. represents the mathematical expectation, Represents the original data, Represents the data that is mapped from the original space to the latent space, and from the latent space to the original space, represents the L2 norm; The binary judgment module uses the following loss function to calculate the difference between real data and synthetic data during training: ; In the formula, Represents the cross entropy function between real data and synthetic data, Represents real data, Represents synthetic data. 5. The sensor data protection method according to claim 1, characterized in that the fourth step of expanding the simulated data space adopts a filter combination method to achieve the expansion of the data space, and at the same time adopts a Bayesian optimization method to find each parameter of the filter combination; Among them, the filter combination method is to combine the original data and the filtered data according to the formula after the frequency and combination ratio are cut off according to the simulated data generated by the generative adversarial network: ; In the formula, the first part on the right side of the equal sign represents a certain proportion of the original data, and the second part represents a certain proportion of the filtered data; represents the proportion of filtered data in the combined data, represents the cutoff frequency of the filter, Represents the original data, Represents the data after filtering. Indicates the proportion of original data in the combined data; the left side of the formula equal sign indicates the result after filtering combination; Use Bayesian optimization to find parameters , including optimizing expressions, fitting models, and collecting functions; Determine the optimization expression, use the Gaussian process as the fitting model, and the probability boost function as the acquisition function: ; In the formula, the right side of the formula equal sign represents the distance between the filtered combined data and the original data, and the left side of the formula equal sign represents the specific value of the distance; represents the dynamic time-adjusted distance calculation function, represents the filtered combined data, Show original data; The interception and replacement of sensor data in the fifth step: implement the sensor monitoring module through Hook, and perform bottom-level interception and replacement on the sensor data transmission interface; find the module class android.hardware.SystemSensorManager that controls the distribution of sensor data in the Android8.0 system source code, find the specific sensor processing subclass SensorEventQueue in the module class, and the distribution function dispatchSensorEvent therein; Hook the dispatchSensorEvent method under SystemSensorManager in the system service process, load the pre-compiled replacement function module, and replace the sensor interface with synthetic data. 6. A computer device, characterized in that the computer device comprises a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the sensor data protection method according to any one of claims 1 to 5. 7. An information data processing terminal, characterized in that the information data processing terminal is used to implement the sensor data protection method described in any one of claims 1 to 5. 8. A sensor data protection system for implementing the sensor data protection method according to any one of claims 1 to 5, characterized in that the sensor data protection system comprises: The system initialization module is used to realize the user input action distribution, build a transfer matrix through the predefined action transition probability and the user input action distribution, and perform multiple rounds of iterations to verify whether a stable distribution can be achieved, providing feasibility support for the subsequent generation of action sequences; The simulated action sequence construction module is used to use the proposed distribution and the accepted distribution of the constructed transfer matrix, combined with the random walk algorithm to generate a behavior action sequence that conforms to the predefined action distribution, providing data support for the subsequent sensor simulated data arrangement rules; The sensor simulation data generation module is used to pre-train the generative adversarial network model with real data, so that the accuracy of the simulation data generated by the model in the action recognition task reaches more than 90%. For each action, multiple sets of data are generated as a buffer to provide the original data template for the subsequent simulation data space expansion task; Expand the simulated data space module to extract the data of each action in the buffer, combine them according to the filter combination rules, and use the Bayesian optimization algorithm to select multiple parameters that can achieve local optimality. Since data with large differences from the original data but the same classification can be generated, this appropriately solves the problem of mode collapse that may exist in the generative adversarial network; The data combination and replacement module is used to generate simulated data according to the behavioral action sequence generated by the sensor simulated data, hook the sensor data distribution interface at the bottom layer of the mobile device, replace the batch of sensor data, and protect the privacy and security of sensor data from the mobile terminal data release link. 9. The sensor data protection system according to claim 8, characterized in that the sensor data protection system further comprises: a generator, a discriminator; The generator includes an embedding function module, a recovery function module, an embedding recovery loss calculation module, a multi-scale circulation module, and a timing function module; The embedding function module is used to map data from a low dimension in the original space to a high dimension in the latent space; the recovery function module is connected to the embedding function module and is used to accurately restore data from a high-dimensional latent space to a low-dimensional real space; the embedding recovery loss calculation module is used to calculate the difference between the real data and the original data after being processed by the embedding function module and the recovery function module, and is used to repeatedly train the embedding function module and the recovery function module so that the original data can be accurately expressed in the high-dimensional space; the multi-scale cycle module is used to learn the time domain characteristics of each dimension of the multi-sensor and the correlation between the time domain characteristics of each dimension; the timing function module is used to better represent the synthetic data output by the generator in the high-dimensional latent space during the adversarial training process; The discriminator includes a binary judgment function module and a similarity calculation module; the binary judgment function module is used to distinguish real data from synthetic data during adversarial training; the similarity calculation module is connected to the binary judgment function module and is used to calculate the cosine similarity between the synthetic data and the real data in the low-dimensional original space; The embedding function module and the restoration function module are both composed of a multi-scale recurrent neural network and a fully connected network layer. The multi-scale recurrent neural network is composed of one-dimensional recurrent neural network layers of different sizes. The output of each node in the last layer of the multi-scale recurrent neural network is used as the input of the fully connected layer. The timing function module includes a fully connected network and a GRU network; The embedding recovery loss calculation module uses the following formula to calculate the difference between the original data and the data processed by the embedding function module and the recovery function module; The binary judgment module uses the following loss function to calculate the difference between real data and synthetic data during training.